ICLR 2026: Training AI Agents to Reason Harder Increases Tool-Call Hallucinations — 96% of Enterprises Already Running Agents

Asanify AI News·medium signal

An ICLR 2026 paper found that training models to reason more intensively causes them to hallucinate more of the tools they should call — a direct risk for agentic systems that rely on accurate tool invocation. Separately, OutSystems' 2026 State of AI Development survey found 96% of enterprises now run AI agents, while Deloitte's research shows 47% of enterprise AI users have based at least one major business decision on hallucinated content. The combination of near-universal agent deployment with structurally worsening tool hallucination in advanced reasoning models creates an underappreciated reliability gap.